fix(cli/serve): import serve handlers lazily so the CLI works without the serve extra

[Anai-Guo]

What this fixes

Fixes the transformers CLI crashing with NameError when the optional serve extra (which provides openai) is not installed.

Importing any CLI command triggers this chain:

transformers <cmd>
 └─ cli/transformers.py  ->  from transformers.cli.serve import Serve
     └─ serve.py          ->  from .serving.utils import set_torch_seed   # triggers serving/__init__
         └─ serving/__init__.py -> from .server import build_server
             └─ server.py -> from .chat_completion import ChatCompletionHandler   # unconditional
                 └─ chat_completion.py:
                       class TransformersCompletionCreateParamsStreaming(CompletionCreateParamsStreaming, ...):

CompletionCreateParamsStreaming (and the other openai.types symbols) are only imported inside if is_serve_available():, but the handler classes reference them at class-definition time. With the serve extra absent, is_serve_available() is False, the symbol is never imported, and the module raises:

NameError: name 'CompletionCreateParamsStreaming' is not defined

Because server.py imports the four handler modules unconditionally (only the fastapi imports were guarded), the whole CLI becomes unusable without serve deps — e.g. transformers env, as reported in #46468.

The fix

In cli/serving/server.py:

1. Add from __future__ import annotations so build_server's parameter/return annotations (ChatCompletionHandler, FastAPI, …) are not evaluated at definition time.
2. Move the four serve-only handler imports (chat_completion, completion, response, transcription) into the existing if is_serve_available(): guard, next to the fastapi imports. ModelManager and .utils stay unconditional — they have no serve-only dependencies.

This is sufficient and minimal:

• build_server only uses the handlers in annotations, never as values in its body, so the guarded imports aren't needed at import time.
• serve.py already imports the handler classes itself (lazily, inside the run path) when it actually starts a server, so guarding them in server.py does not affect serving.
• tests/cli/test_serve.py imports only build_server from this module (and is gated with @require_serve), so it is unaffected.

Verification

With the serve extra absent (is_serve_available() is False):

• Before: python -c "import transformers.cli.serving.server" → NameError: name 'CompletionCreateParamsStreaming' is not defined
• After: imports cleanly; transformers env / version / download work again.

With the extra present, serving behaviour is unchanged.

Note: the TypeError in the issue's main traceback originates in the model's remote modeling_nemotron_h.py (trust_remote_code), which is outside this repo; this PR addresses the transformers env NameError regression reported in the same issue.

🤖 Generated with Claude Code

[Rocketknight1]

cc @LysandreJik for serve/CLI final approval but the fix LGTM